runtime: make mutexActiveSpinSize configurable

[weijinrongyuan]

Proposal Details

Currently, in src/runtime/lock_spinbit.go, the constant mutexActiveSpinSize is hardcoded to 30，My tests on ARM64 platform indicate that the hardcoded value of 30 is not optimal。Whether the mutexActiveSpinSize value can be made configurable？

[seankhliao]

Do you have benchmark results of how far from optimal it is?
Is a good value something that can be detected rather than made user configurable?

[dmitshur]

CC @golang/runtime.

[weijinrongyuan]

Do you have benchmark results of how far from optimal it is? Is a good value something that can be detected rather than made user configurable?
I ran a subset of the Go bent benchmark suite on an ARM64 platform，I changed the hardcoded mutexActiveSpinSize from 30 to 40。The results: 5%–15% throughput improvement in 8 out of 10 benchmarks.

[rhysh]

I agree that "30" is unlikely to be the best value for GOARCH=arm64.

I expect the path forward will be to find a new value (or new design) for those delay loops. It's possible that auto-detection could work, if we can't find a single good value/design. I think it's unlikely that we'd make it outright configurable.

When I last looked into this, I wrote #68578 (comment) and opened #69232. The approach in #69232 (comment) / https://developer.arm.com/community/arm-community-blogs/b/architectures-and-processors-blog/posts/multi-threaded-applications-arm of "pick a desired duration, and loop in a particular way until that much time has passed" seems promising, and would allow better sharing of the design across architectures (and processor generations).